Category: Max’s Reports

After meeting with my team, we have decided to leave the ML integration into the system as a later issue that may not become fully realized due to my ongoing health issues. I have been setback, but after reorganizing our groups schedule, I have reformed my goals for this week to be finishing out the current base ML model. I am still working on it, and am therefore on schedule for the new schedule, but have also been working to mitigate any risks with incorporating it with the Jetson by working on the Jetson Nano and starting setup when training my model.

This week I was able to continue work on the transfer layers, however, due to some serious health complications, I was fairly limited in my time to work on this. Due to this setback, our team is looking into integration alternatives as we look to start integration this weekend. Aside from this, though setback due to health, I am only slightly behind schedule.

This week I was finally able to produce some results for the simplified ML classifier, and assisted my team in making the decision to go forward with implementing the Jetson over the Raspberry Pi. I have moved onto finalizing the additional transfer layers that will be used to create our final classifier, which is on schedule. I have also been doing more research on potential for parallelization through NVIDIA options available through the Jetson. This puts me back on schedule. In addition, the NVIDIA research is to help avoid any complications that may arise from implementing the classifier as I am, and provide shortcuts that will enable me to speedup the classifier if we run into future problems with our speed. All together, this puts me on schedule.

This week I have been working on a simplified version of our final ML classifier in order to start performing preliminary speed and efficiency tests for the CNN. Our group is still bouncing around a few options in terms of where we are going to host the CNN, whether its on a Pi, Jetson, or externally. Since we are still deciding, this has put some extra stress on my portion of the project to produce some testing results, so although I was originally on track, I am a little behind due to the decision our team has made to explore potentially different implementations of our system. However, I am now much more accustomed to InceptionV4 and Tensor-flow, and have a better understanding of the transfer layers I will be implementing, but have focused on creating a classifier without extensive transfer layers in order to start platform testing.

Our group was finishing up our design report this week. However, since our design choices and implementation for my ML portion of the project have not changed much since the start, while I was primarily finishing up the design report with my teammates, I was also able to continue to implement my InceptionV4 architecture.

As for schedule, I am on track with where I wanted to be. This week I continued my work on the InceptionV4 architecture and I am on track for that task of making a simple classifier. I was also able to better flesh out the additional transfer layers that will be required.

Our group is finishing up some final design implementation choices. The current options do not effect me, so I have been free to continue my work without and hardware hiccups due to changing our main hardware. As such, this week has been primarily implementing the InceptionV4 architecture and setting up a functional dog/cat breed classifier. I am still working on this, but the work is on track.

As for schedule, I am back on track with where I wanted to be. Primary research is completed and implementation is starting. This week I was able to start implementing the InceptionV4 architecture and I am on track for that task.

Our group finally settled on using a Raspberry Pi rather than an FPGA or a Jetson as our hardware. This has meant I no longer need to research neural network implementation on FPGA, but we did receive useful advice in our weekly meeting as a result of our team’s research into this topic. However, this has given me more time to research the exact nature of our neural network. We are definitely going to be moving forward with a CNN. As for what extra layers we will be using on top of our architecture, this is still being determined and will most likely have to be fine tuned over the course of the semester. However, it currently looks like our transfer learning model will implement an initial convolutional layer, followed by 5 to 6 feature layers that will be trained via the user uploaded images of their pet(s).

As for schedule, I am back on track with where I wanted to be. Primary research is completed and implementation is starting. This week I want to start to get a functional transfer learning model as a starting point for testing and training.

Neural networks were covered in 15-301, Introduction to Machine Learning, which I took last year. I have done extra study on various forums looking at examples using the Inception architecture to become more familiar with this particular piece of additional software, but I am already comfortable with the area due to Intro to ML and Computational Neuroscience, which also covered neural networks.

Our group has had meetings to further flesh out our design implementation. One main topic that came up was the possibility of using an FPGA over a Raspberry Pi. This was inspired by witnessing other presentations that were performing computation on live video (rotoscoping, multiple object tracking) that is similar to our projects in terms of the level of computation on live video. Due to this, I have been researching the viability of an FPGA in our solution to improve classification speed of our classifier while still allowing easy video feed input to the FPGA and the ability to communicate with the web application. In addition, I have been researching existing cat/dog breed classifiers and working with InceptionV3 (our currently chosen ML architecture) to see if the current architecture will work.

Due to some major personal setbacks this week I am behind schedule, specifically on evaluating our current ML architecture choice. In addition, as of writing this, I am still unsure as to the viability of using an FPGA and am meeting with my group to further discuss this design change. The main priority for this week is to finish our design review presentation, which will first require a final decision as to whether or not we move forward with an FPGA, Raspberry Pi, or something else as our hardware. In addition, a complete evaluation of the InceptionV3 architecture as it compares to other existing architectures is to be completed soon. This should kick off work on implementing a animal classifier on the chosen architecture, which should be available this week.